Predicting the diffusion coefficients of rejuvenators into bitumens using molecular dynamics, machine learning, and force field atom types

Abstract

This study explores the use of chemical descriptors derived from force field atom types to predict Fickian diffusion coefficients of rejuvenators in bitumen, utilizing machine learning models trained on data from 240 non-equilibrium molecular dynamics simulations. The simulations cover three bitumen types (NO, TO, FO), five aging degrees, and four temperatures (60 °C, 120 °C, 160 °C, 200 °C), capturing diffusion coefficients ranging from 0.0068e-10 m²/s in highly aged bitumens at 60 °C to 4.35e-10 m²/s in fresher samples at 200 °C. The MLM, built with 18 chemical descriptors for bitumen and rejuvenator sides, achieves an R² of 0.97, accurately predicting diffusion across varied conditions. This approach abstracts away from the need for repeated MD simulations, enabling diffusion predictions even for systems outside the original dataset. The manuscript presents three case studies to illustrate how the model can be used for the iterative design of rejuvenators by optimizing molecular structures based on critical chemical features, such as rejuvenator oxygen content, bitumen sulfur content, and molecular weights. It also demonstrates how the model offers a practical framework for understanding the diffusion and performance of rejuvenators by linking time-dependent factors—such as concentration, depth, and rejuvenation time—with the bulk properties of bitumen-rejuvenator systems, facilitating industrial applications.